All Leaks Count, Some Count More: Influence under re-prompting is not internal provenance
Shapley-DCLR attributes influence to visible rationale claims by re-prompting on subsets of them. That measures the effect of a later prompt, not what caused the original answer.
The metric assigns Shapley values to the claims in a visible rationale by evaluating subsets of those claims, then weights their leakage labels by absolute Shapley magnitude, reporting the fraction of decision-driving reasoning that is post-cutoff. The arithmetic is a legitimate attribution construction for the chosen set function.
The intervention changes a later prompt. It does not observe which information produced the original answer. A model can recall the outcome straight from its weights and be insensitive to every rationale claim, in which case all marginal contributions are zero while the leakage is total. The converse also happens: explicit claims can move the follow-up answer without faithfully describing the original reasoning.
So a low score is not evidence of clean reasoning, and the fraction is not identified as the fraction of internal reasoning caused by future knowledge.
As a measure of how much a stated rationale drives a re-prompted answer, the construction is sound and the weighting by decision impact is a genuine improvement on counting leaked claims equally. TimeSPEC's date-filtered retrieval does control the retrieved records. A closed-world instruction does not isolate the model from its parametric memory.
Shapley-DCLR measures claim influence under subset re-prompting, not the causal provenance of the original answer. Hidden memory can drive an answer while every visible claim has zero marginal influence. The score should not be read as a certified fraction of contaminated reasoning.